Designing Incentive Compatible Payment Rules for Combinatorial Auctions with Structural SVMs

Combinatorial auctions have a wide range of real-world applications; yet, designing combinatorial auction mechanisms that simultaneously possess good economic properties and computational tractability remains a major challenge. An auction mechanism consists of an allocation rule and a payment rule. We propose a new framework that uses Structural SVMs to design a payment rule for any given allocation rule. Besides being tractable, the payment rule produced by an exact classifier is both strategyproof and individually rational. Unlike the VCG payment rule, our framework does not require an optimal allocation rule, an NP-hard problem, in order to obtain a strategyproof mechanism. Our experiments show that the payment rules generated from our framework enjoy a low level of ex post regret and approximate well-known strategyproof payment rules, such as VCG and second price, reasonably well. In addition, applying our framework to an allocation rule with no corresponding strategyproof payment rule does not result in additional performance loss.